Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

False-pass bias in LLM judges silently disables skill retirement past a sharp threshold that more data cannot fix.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-10 19:00 UTC pith:FKW2M5S7

load-bearing objection Solid behavioral-safety result: false-pass bias (not noise) kills contribution-based retirement past a sharp, data-irrecoverable cliff, often silently; the exogenous-channel scope is real but already declared. the 3 major comments →

arxiv 2607.07436 v1 pith:FKW2M5S7 submitted 2026-07-08 cs.AI cs.CLcs.CR

The Blind Curator: How a Biased Judge Silently Disables Skill Retirement in Self-Evolving Agents

Xing Zhang , Yanwei Cui , Guanghui Wang , Ziyuan Li , Wei Qiu , Bing Zhu , Peiyang He This is my paper
classification cs.AI cs.CLcs.CR
keywords self-evolving agentsskill retirementLLM-as-judgefalse-pass biaslibrary driftcorrupted rewarddefect-injection auditRatchet
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Self-evolving agents keep a skill library healthy by retiring skills that stop helping, but that governance assumes an honest signal of which skills failed. On reference-free tasks the only practical grader is an LLM judge, and its errors are not white noise: they tend to report true failures as passes. This paper shows that such false-pass bias does not merely add noise; past a sharp threshold it turns off contribution-based retirement entirely, at any sample size. Symmetric noise leaves the curator working. The mechanism failure is universal across report-writing and code-generation testbeds and across scarce and abundant failures, while aggregate task quality often stays flat, so the disabled curator is silent. The practical payoff is a cheap defect-injection audit that places any judge on either side of the threshold before self-evolution is trusted.

Core claim

Under the modeled binary channel, false-pass bias displaces the observed pass rate of skills so that contribution-based retirement stops firing once the false-pass rate reaches (1−τ)/2; beyond that cliff no amount of data restores retirement. Symmetric noise only compresses the signal and leaves retirement intact. Separating genuine contribution-retirement from cap-eviction churn shows the mechanism failure holds across domains and failure rates, while outcome harm appears only when the same corruption also starves skill synthesis—so the curator can die without any aggregate metric moving.

What carries the argument

The bias-aware non-divergence bound (Proposition 1′) on the corrupted channel: observed pass rate concentrates on κp̄(s)+ρF→P, so false-pass bias shifts the retirement statistic upward past an unrecoverable threshold while symmetric noise only attenuates it.

Load-bearing premise

The judge’s false-pass rate is treated as an exogenous, skill-independent constant that can be measured offline; the paper does not cover a learned judge whose blindness the skill library itself could learn to exploit.

What would settle it

Run the same evolution loop with a judge whose measured false-pass rate sits above (1−τ)/2 and still observe non-zero genuine contribution-retirement of bad skills at large trial counts; or show that a live learned judge’s false-pass rate stays safely below the cliff for every skill the library invents.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper studies how asymmetric reward corruption affects skill retirement in failure-driven self-evolving agents (built on the authors’ Ratchet governance stack). Modeling an LLM judge as a binary channel, it derives a bias-aware non-divergence bound (Prop. 1′) and a sharp cliff: contribution-based retirement becomes impossible for any sample size once the false-pass rate satisfies ρ_F→P ≥ (1−τ)/2, whereas symmetric noise only attenuates the signal. Empirically, the authors inject exogenous corruption on top of a deterministic QC grader in a reference-free report-writing testbed (with an MBPP+ code-generation cross-check), separate genuine contribution-retirement from cap-eviction churn, and show that false-pass bias drives true retirement to ~0 past the cliff across subsets and domains, while outcome harm is regime-dependent and often silent in aggregate eval. A defect-injection audit is proposed as a pre-deployment go/no-go test for a judge’s false-pass rate.

Significance. If the result holds within its stated scope, this is a useful behavioral-safety contribution for self-evolving agents: it identifies a concrete, falsifiable failure mode (curator blindness under false-pass bias), gives an operational threshold in terms of τ, and supplies a cheap offline audit rather than another end-to-end performance claim. Strengths include (i) a clean channel algebra with an explicit cliff prediction, (ii) causal isolation via exogenous corruption on training labels only while evaluating with the true grader, (iii) honest decomposition of genuine retirement vs. cap-eviction, (iv) cross-domain replication including a verifier-backed code setting, and (v) explicit ownership that the contribution is mechanistic safety, not lift. The exogenous, skill-independent corruption assumption is the main external threat to transfer, and is already flagged in Appendix C and §9.

major comments (3)
  1. Abstract, Fig. 5, and §8 playbook: the claim that a one-shot defect-injection audit tells an operator which side of the cliff their judge occupies is load-bearing for the practical contribution, but it is only guaranteed under the exogenous, skill-independent channel of Prop. 1′ / Appendix C. The adversarial-coupling remark (Appendix C) and §9 correctly note that a learned judge whose blindness co-evolves with skill phrasing can make ρ_F→P skill-dependent and unbounded by a fixed audit. Please state this scope limit in the abstract and playbook (not only in Limitations), and qualify the go/no-go language accordingly—e.g., the audit bounds a static judge, not a library that can learn to exploit the judge.
  2. §6 Table 1 and §7: on REPORT-MAIN-71 the clean loop already sits below the no-skill floor (Δ = −0.060), so the “harm ranking” is relative to a non-improving baseline. BAND-58 and HARD-133 partially address headroom, but the paper still reports no detectable end-to-end lift under a paired continuous re-measurement (Appendix G). For a safety paper this is acceptable, but the framing in §1 and the conclusion should more sharply separate (a) the universal mechanism claim (true-ret → 0), which the data support, from (b) any implication that operators lose measurable quality when the curator dies in abundant-failure regimes—Fig. 4c shows they often do not. Tighten the abstract’s “silently switches off” language so silence is not read as hidden quality loss in every regime.
  3. §4 / Appendix C, pure false-pass case (ρ_P→F = 0): experiments inject pure false-pass bias, while the audited real LLM judge sits in the opposite corner (ρ_F→P ≈ 0.01, ρ_P→F ≈ 0.95). The theory covers mixed channels via κ and ρ_F→P, but the empirical “lenient judge” region that hits the cliff is only synthetic. Please add at least one mixed-channel or deliberately lenient-judge condition (or a clear negative result if unavailable) so the cliff is not demonstrated solely by synthetic pure-q sweeps; otherwise state that the dangerous operating point is extrapolated from the channel model rather than observed in a live LLM judge.
minor comments (6)
  1. §3–§4: π_τ := (1−τ)/2 is introduced cleanly, but q and ρ_F→P are used interchangeably; a single notation pass would help readers tracking the cliff.
  2. Fig. 3a vs Table 1: at q=0.2 true retirement is already 0 on MAIN-71; a short note that this is consistent with few skills having p̄ ≤ (π_τ−q)/(1−q) under a strong composer would prevent readers from thinking the empirical cliff is softer (or harder) than theory.
  3. Fig. 8 is useful but hard to read in grayscale; consider a line plot of cumulative genuine retirements over rounds as a companion.
  4. Related work: the class-conditional label-noise analogy is apt; a brief pointer to standard asymmetric-noise results (beyond the catastrophic-forgetting analogy) would situate Prop. 1′ for readers outside agent skill libraries.
  5. Typos / polish: “itsilently” (Abstract/Intro), “mech-anism” line break, and “gradual” vs “graceful” wording around noise could be cleaned in a final pass.
  6. Table 2 is excellent for reproducibility; if code/data release is planned, state it explicitly in §10 or a reproducibility statement.

Circularity Check

1 steps flagged

No derivation circularity: the false-pass cliff is algebra from the channel model, and experiments inject exogenous corruption; only mild self-citation of the authors' Ratchet setup.

specific steps
  1. self citation load bearing [Sec. 2 'The Ratchet mechanism we build on'; Sec. 4; Prop. 1′ / Appendix C]
    "We start from Ratchet (Zhang et al., 2026a;b), a minimal recipe on a frozen LLM: ... Crucially, a Curator retires any skill whose empirical contribution falls below a threshold after enough trials, under a hard cap on active skills. Retirement-plus-cap yields a non-divergence guarantee... We adopt Ratchet because it is, to our knowledge, the only self-evolving-skill scheme with such a guarantee. That makes it the natural object on which to ask when the guarantee survives an imperfect reward, since its proof is the thing an unreliable judge can break."

    The non-divergence object and retirement rule being stress-tested are justified by the authors' own prior Ratchet papers, including a mild uniqueness claim. This is load-bearing for *which* curator is studied and for the baseline Prop. 1 that Prop. 1′ extends, but it does not force the false-pass cliff or the empirical mechanism result: those follow from new channel algebra and exogenous corruption experiments independent of the Ratchet guarantee's truth.

full rationale

The load-bearing theoretical step is the corrupted-reward channel p_obs = κ p̄ + ρ_F→P and the retirement rule p̂ ≤ π_τ = (1−τ)/2, which together imply that false-pass bias disables retirement once ρ_F→P ≥ (1−τ)/2 at any sample size (Sec. 4, Prop. 1′, Appendix C). That implication is ordinary algebra from stated definitions, not a fit renamed as a prediction and not self-definitional. Empirically, corruption is injected exogenously on a deterministic grader at rates that bracket the predicted cliff (q ∈ {0.2, 0.45, 0.7}), genuine contribution-retirement is separated from cap-eviction, and the signature is checked across subsets and MBPP+; none of these steps reduce the outcome to the input by construction. The only circularity-adjacent element is that the object of study (Ratchet retirement-plus-cap and its unbiased non-divergence guarantee) is the authors' own prior work. That self-citation supplies the setup and the baseline Prop. 1 being extended, but it does not force the bias cliff, the inverted-U outcome pattern, or the silent-curator claim—those are new channel analysis and independent measurements. Under the stated rules this is minor, non-load-bearing self-citation of prior mechanism work, not a circular derivation of the central result. Score 2.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 1 invented entities

The central claim rests on the Ratchet retirement rule and non-divergence setup (authors’ prior work), a binary class-conditional corruption channel, exogenous injection of that channel on a deterministic QC grader treated as ground truth, and fixed governance hyperparameters. No new physical entities are postulated; the ‘blind curator’ names an observed process failure. Free parameters are the retirement threshold and sample/cap settings that locate the cliff and run the loop.

free parameters (4)
  • retirement threshold τ
    Sets the cliff location π_τ = (1−τ)/2; experiments fix τ=0.10 so the predicted cliff is at 0.45. Chosen by the governance recipe, not fitted to the bias result, but the quantitative threshold depends on it.
  • Nmin (min trials before retirement)
    Fixed at 24; required sample size for the Hoeffding concentration in Prop. 1′. Affects when retirement can fire but not the asymptotic cliff under false-pass bias.
  • skill bank cap C
    Fixed at 12; creates cap-eviction churn that must be separated from genuine contribution-retirement. Central to the measurement design.
  • corruption rates ρ and q
    Swept experimentally (ρ∈{0.1..0.4}, q∈{0.2,0.45,0.7}); not fitted to claim a continuous curve but chosen to bracket the predicted cliff.
axioms (6)
  • domain assumption Ratchet retirement rule: retire skill s when observed contribution ĉ(s) ≤ −τ after n(s) ≥ Nmin (equivalently observed pass rate ≤ (1−τ)/2).
    Imported from the authors’ prior Ratchet papers; the entire bias analysis is relative to this rule (§1, §4, Appendix C).
  • domain assumption Binary corruption channel: Pr[ỹ=1|y=0]=ρ_F→P, Pr[ỹ=0|y=1]=ρ_P→F, with observed pass rate concentrating on κ p̄(s) + ρ_F→P.
    Standard class-conditional label-noise model specialized to pass/fail; load-bearing for the cliff derivation (Appendix C).
  • domain assumption Failures are the single shared input to both synthesis and retirement in the evolution loop.
    Structural premise that makes false-pass bias damage both stages at the source (§3).
  • ad hoc to paper Deterministic QC checks (orphan citation, unregistered metric, bare number, broken xref, missing TL;DR) are treated as ground-truth y for mechanism isolation.
    Enables clean corruption injection on a reference-free task; validity is scoped to citation-discipline QC, not insight (§3, §9).
  • domain assumption Router and cap conditions of original Ratchet Prop. 1 hold so the bias-aware floor (Prop. 1′) applies.
    Stated in Appendix C proof sketch; not re-proved here.
  • ad hoc to paper Corruption is exogenous and independent of skill content (no adversarial coupling).
    Explicit experimental design choice; learned-judge coupling declared out of scope (Appendix C remark, §9).
invented entities (1)
  • blind curator failure mode independent evidence
    purpose: Name the process-level failure in which false-pass bias breaks Judge→Curator evidence so contribution-retirement stops while aggregate outcomes can look normal.
    Descriptive label for the observed mechanism, not a new physical or mathematical object; independent handle is the measurable ρ_F→P vs cliff comparison.

pith-pipeline@v1.1.0-grok45 · 22104 in / 3829 out tokens · 45580 ms · 2026-07-10T19:00:05.361894+00:00 · methodology

0 comments
read the original abstract

A self-evolving agent retires its bad skills by watching them fail, so what happens when the judge cannot see the failures? Skill retirement is the structural constraint that keeps a growing library from drifting below the no-skill baseline, but its guarantee assumes an unbiased reward, which is false for the LLM judges that reference-free tasks force upon us. We show that a biased judge does not merely add noise; it \emph{silently switches off the curator}. We make this precise with a corrupted-reward analysis and, isolating the causal channel by injecting corruption on top of a deterministic reward, a behavioral study on a reference-free report-writing testbed with a code-generation cross-check. Symmetric noise leaves retirement intact, but \emph{false-pass} bias (failures slipping through as passes) disables contribution-based retirement past a sharp threshold that no amount of data can cross. Separating genuine retirement from cap-eviction churn shows this \emph{mechanism} failure is universal, holding across domains and failure rates and sparing only near-zero-false-pass, verifier-like graders. The downstream \emph{outcome}, though, is regime-dependent: eval quality degrades only where the same corruption also starves skill synthesis, and otherwise holds steady, so the disabled curator is \emph{silent}, surfacing in no aggregate metric. The contribution is a behavioral safety result, not a performance one. A cheap defect-injection audit then tells an operator, before deployment, which side of the threshold their judge occupies.

Figures

Figures reproduced from arXiv: 2607.07436 by Bing Zhu, Guanghui Wang, Peiyang He, Wei Qiu, Xing Zhang, Yanwei Cui, Ziyuan Li.

Figure 1
Figure 1. Figure 1: The blind curator failure mode. The same failure-driven loop (solve, judge, re￾tire) under an honest reward (left) and a false-pass judge (right): false passes break the Judge→Curator evidence, so retirement quietly stops while aggregate outcomes can look normal. The gap opens at a sharp threshold. We treat that asymmetry as the object of study. Our thesis turns on two separable knobs of the reward channel… view at source ↗
Figure 2
Figure 2. Figure 2: Flag rate on injected defects (green = deterministic QC; purple = held-out LLM [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: REPORT-MAIN-71 (Claude Haiku 4.5, 3 seeds), as a fraction of the clean-reward level. (a) Mechanism. Observed failures are the loop’s single input; rising false-pass bias q shrinks that pool and starves both downstream stages, synthesis and genuine retirement (which hits zero past the cliff). (b) Outcome. Eval damage vs the clean-reward loop, against realised corruption (the fraction of training labels actu… view at source ↗
Figure 4
Figure 4. Figure 4: The causal chain behind silent curator failure, across three subsets: REPORT-BAND￾58 (scarce failures), REPORT-HARD-133 (abundant, same domain), and MBPP+ HARD100 (abundant, different domain and model). (a) genuine retirement falls to ≈0 past the cliff in every subset, the curator dies universally. (b) synthesis survives where failures stay abundant and starves only in the scarce subset. (c) so the eval ou… view at source ↗
Figure 5
Figure 5. Figure 5: Deployment playbook, assembling the paper’s findings into a before-deployment [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: The retirement signature replicates across two subsets (3 seeds; blue = noise, red = [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: A real claim-negation defect (demo-credo): flipping the direction word (“grew” → “fell”) leaves the number (205.68%), metric tag {m_rev_yoy}, and citation [E0001] byte-identical, so deterministic QC (and even a value check) sees nothing wrong, yet the sentence now contradicts its own cited source. The QC-invisible / judge-visible gap in one example. 0 1 2 3 4 5 6 7 8 9 10 11 round clean QC noise = 0.1 nois… view at source ↗
Figure 8
Figure 8. Figure 8: Per-round bank deprecations (mean of 3 seeds), [PITH_FULL_IMAGE:figures/full_fig_p015_8.png] view at source ↗

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. When the Reward Suite Is Leaky: A Preregistered Causal Contrast of Natural Verifier False Positives in RLVR

    cs.LG 2026-07 accept novelty 7.0

    Natural false positives in MBPP rewards inflate train reward and pay for real bugs, yet a preregistered leaky-vs-hardened GRPO contrast finds held-out non-inferiority and selection of pre-existing errors, not learned ...

Reference graph

Works this paper leans on

20 extracted references · 20 canonical work pages · cited by 1 Pith paper · 17 internal anchors

  1. [1]

    Program Synthesis with Large Language Models

    Jacob Austin, Augustus Odena, Maxwell Nye, Maarten Bosma, Henryk Michalewski, David Dohan, Ellen Jiang, Carrie Cai, Michael Terry, Quoc Le, and Charles Sutton. Program synthesis with large language models.arXiv preprint arXiv:2108.07732,

    work page internal anchor Pith review Pith/arXiv arXiv

  2. [2]

    Constitutional AI: Harmlessness from AI Feedback

    Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, et al. Constitu- tional AI: Harmlessness from AI feedback.arXiv preprint arXiv:2212.08073,

    work page internal anchor Pith review Pith/arXiv arXiv

  3. [3]

    Closing the Feedback Loop: From Experience Extraction to Insight Governance in Verbal Reinforcement Learning

    Yanwei Cui, Xing Zhang, Yulong Zhang, Li Shao, Xiaofeng Shi, Guanghui Wang, and Peiyang He. Closing the feedback loop: From experience extraction to insight governance in verbal reinforcement learning.arXiv preprint arXiv:2606.17591,

    work page internal anchor Pith review Pith/arXiv arXiv

  4. [4]

    Under review

    9 Preprint. Under review. Xu Huang, Junwu Chen, Yuxing Fei, Zhuohan Li, Philippe Schwaller, and Gerbrand Ceder. CASCADE: Cumulative agentic skill creation through autonomous development and evolution.arXiv preprint arXiv:2512.23880,

    work page arXiv

  5. [5]

    LLMs-as-Judges: A Comprehensive Survey on LLM-based Evaluation Methods

    Haitao Li, Qian Dong, Junjie Chen, Huixue Su, Yujia Zhou, Qingyao Ai, Ziyi Ye, and Yiqun Liu. LLMs-as-judges: A comprehensive survey on LLM-based evaluation methods.arXiv preprint arXiv:2412.05579,

    work page internal anchor Pith review Pith/arXiv arXiv

  6. [6]

    SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks

    Xiangyi Li, Wenbo Chen, Yimin Liu, Shenghan Zheng, Xiaokun Chen, Yifeng He, Yubo Li, Bingran You, Haotian Shen, Jiankai Sun, et al. SkillsBench: Benchmarking how well agent skills work across diverse tasks.arXiv preprint arXiv:2602.12670,

    work page internal anchor Pith review Pith/arXiv arXiv

  7. [7]

    SkillsVote: Lifecycle Governance of Agent Skills from Collection, Recommendation to Evolution

    Hongyi Liu, Haoyan Yang, Tao Jiang, Bo Tang, Feiyu Xiong, and Zhiyu Li. SkillsVote: Lifecycle governance of agent skills from collection, recommendation to evolution.arXiv preprint arXiv:2605.18401,

    work page internal anchor Pith review Pith/arXiv arXiv

  8. [8]

    Trace2Skill: Distill Trajectory-Local Lessons into Transferable Agent Skills

    Jingwei Ni, Yihao Liu, Xinpeng Liu, Yutao Sun, Mengyu Zhou, Pengyu Cheng, Dexin Wang, Xiaoxi Jiang, and Guanjun Jiang. Trace2Skill: Parallel inductive skill distillation for LLM agents.arXiv preprint arXiv:2603.25158,

    work page internal anchor Pith review Pith/arXiv arXiv

  9. [9]

    MemGPT: Towards LLMs as Operating Systems

    Charles Packer, Sarah Wooders, Kevin Lin, Vivian Fang, Shishir G. Patil, Ion Stoica, and Joseph E. Gonzalez. MemGPT: Towards LLMs as operating systems.arXiv preprint arXiv:2310.08560,

    work page internal anchor Pith review Pith/arXiv arXiv

  10. [10]

    AutoPyVerifier: Learning Compact Executable Verifiers for Large Language Model Outputs

    Pouya Pezeshkpour and Estevam Hruschka. AutoPyVerifier: Learning compact executable verifiers for large language model outputs.arXiv preprint arXiv:2604.22937,

    work page internal anchor Pith review Pith/arXiv arXiv

  11. [11]

    Dynamic Skill Lifecycle Management for Agentic Reinforcement Learning

    Junhao Shen, Teng Zhang, Xiaoyan Zhao, and Hong Cheng. Dynamic skill lifecycle man- agement for agentic reinforcement learning.arXiv preprint arXiv:2605.10923,

    work page internal anchor Pith review Pith/arXiv arXiv

  12. [12]

    Large Language Models are Inconsistent and Biased Evaluators

    Rickard Stureborg, Dimitris Alikaniotis, and Yoshi Suhara. Large language models are inconsistent and biased evaluators.arXiv preprint arXiv:2405.01724,

    work page internal anchor Pith review Pith/arXiv arXiv

  13. [13]

    Voyager: An Open-Ended Embodied Agent with Large Language Models

    Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. Voyager: An open-ended embodied agent with large language models.arXiv preprint arXiv:2305.16291,

    work page internal anchor Pith review Pith/arXiv arXiv

  14. [14]

    Agent Workflow Memory

    Peiyi Wang, Lei Li, Liang Chen, Zefan Cai, Dawei Zhu, Binghuai Lin, Yunbo Cao, Lingpeng Kong, Qi Liu, Tianyu Liu, et al. Large language models are not fair evaluators. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 9440–9450, 2024a. 10 Preprint. Under review. Zora Zhiruo Wang, Jiayua...

    work page internal anchor Pith review Pith/arXiv arXiv

  15. [15]

    SkillRL: Evolving Agents via Recursive Skill-Augmented Reinforcement Learning

    Peng Xia, Jianwen Chen, Hanyang Wang, Jiaqi Liu, Kaide Zeng, Yu Wang, Siwei Han, Yiyang Zhou, Xujiang Zhao, Haifeng Chen, et al. SkillRL: Evolving agents via recursive skill-augmented reinforcement learning.arXiv preprint arXiv:2602.08234,

    work page internal anchor Pith review Pith/arXiv arXiv

  16. [16]

    AutoSkill: Experience-driven lifelong learning via skill self-evolution.arXiv preprint arXiv:2603.01145,

    Yutao Yang, Junsong Li, Qianjun Pan, Bihao Zhan, Yuxuan Cai, Lin Du, Jie Zhou, Kai Chen, Qin Chen, Xin Li, et al. AutoSkill: Experience-driven lifelong learning via skill self-evolution.arXiv preprint arXiv:2603.01145,

    work page arXiv

  17. [17]

    SkillAdaptor: Self-Adapting Skills for LLM Agents from Trajectories

    Zhuoyun Yu, Xin Xie, Wuguannan Yao, Chenxi Wang, Lei Liang, Xiang Qi, and Shumin Deng. SkillAdaptor: Self-adapting skills for LLM agents from trajectories.arXiv preprint arXiv:2606.01311,

    work page internal anchor Pith review Pith/arXiv arXiv

  18. [18]

    TextGrad: Automatic "Differentiation" via Text

    Mert Yuksekgonul, Federico Bianchi, Joseph Boen, Sheng Liu, Zhi Huang, Carlos Guestrin, and James Zou. TextGrad: Automatic “differentiation” via text.arXiv preprint arXiv:2406.07496,

    work page internal anchor Pith review Pith/arXiv arXiv

  19. [19]

    Library Drift: Diagnosing and Fixing a Silent Failure Mode in Self-Evolving LLM Skill Libraries

    Xing Zhang, Yanwei Cui, Guanghui Wang, Ziyuan Li, Wei Qiu, Bing Zhu, and Peiyang He. Library drift: Diagnosing and fixing a silent failure mode in self-evolving LLM skill libraries.arXiv preprint arXiv:2605.19576, 2026a. Xing Zhang, Yanwei Cui, Guanghui Wang, Ziyuan Li, Wei Qiu, Bing Zhu, and Peiyang He. Ratchet: A minimal hygiene recipe for self-evolving...

    work page internal anchor Pith review Pith/arXiv arXiv

  20. [20]

    section":

    A Evidence slice: schema and example A trial gives the composer oneevidence slice: the section brief plus the closed set of evidence cards, metric tags, and allowed cross-references it may use. The composer must cite only these cards, annotate every numeral with a card or a registered metric tag, hedge any card marked status: weak , and meet the structura...

    work page 2026